Active Invariant Causal Prediction: Experiment Selection Through Stability
summary
Active Invariant Causal Prediction (called Active ICP) selects interventions to reveal direct parents of Y by testing which conditioning sets remain invariant across environments and returning the intersection; it guarantees S_hat ⊆ S in population but is conservative in finite samples. Practical strategies tested are the empty set strategy, a ratio statistic over invariant sets, and using estimated Markov blankets; simulations (P≈11, samples per intervention 10/100/1000) show the empty set rule is most robust, ABCD gets early power but incurs false positives.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:28Causal structure learning Setting
- 1:04Causal models generalize Invariant models are potentially causal
- 14:23Active causal learning Setting
- 15:06Invariant Causal Prediction (ICP) Example
- 18:00Invariant Causal Prediction (ICP) Heterogeneity plays a key role
- 20:53Intervention selection strategies Some proposals
- 24:55Experiments
- 30:24Active Invariant Causal Prediction
related talks
01h 02mClassical Strategies and Concepts in Causal Discovery
33mLearning Abstract Causal Structure From Causal Interventions In Children And Artificial Agents
28mLearning Causal Representations From Unknown Interventions
24mA Friendly Introduction To Causality
33mAdaptive Prediction Intervals
31mPython and the Holy Grail of Causal Inference
57mA New Perspective on High-Dimensional Causal Inference
31mLearning And Testing Causal Models: A Property Testing Viewpoint